xref: /linux/arch/x86/mm/init.c (revision 4f3c8320c78cdd11c8fdd23c33787407f719322e)
1 #include <linux/gfp.h>
2 #include <linux/initrd.h>
3 #include <linux/ioport.h>
4 #include <linux/swap.h>
5 #include <linux/memblock.h>
6 #include <linux/swapfile.h>
7 #include <linux/swapops.h>
8 #include <linux/kmemleak.h>
9 #include <linux/sched/task.h>
10 
11 #include <asm/set_memory.h>
12 #include <asm/e820/api.h>
13 #include <asm/init.h>
14 #include <asm/page.h>
15 #include <asm/page_types.h>
16 #include <asm/sections.h>
17 #include <asm/setup.h>
18 #include <asm/tlbflush.h>
19 #include <asm/tlb.h>
20 #include <asm/proto.h>
21 #include <asm/dma.h>		/* for MAX_DMA_PFN */
22 #include <asm/microcode.h>
23 #include <asm/kaslr.h>
24 #include <asm/hypervisor.h>
25 #include <asm/cpufeature.h>
26 #include <asm/pti.h>
27 #include <asm/text-patching.h>
28 #include <asm/memtype.h>
29 
30 /*
31  * We need to define the tracepoints somewhere, and tlb.c
32  * is only compied when SMP=y.
33  */
34 #define CREATE_TRACE_POINTS
35 #include <trace/events/tlb.h>
36 
37 #include "mm_internal.h"
38 
39 /*
40  * Tables translating between page_cache_type_t and pte encoding.
41  *
42  * The default values are defined statically as minimal supported mode;
43  * WC and WT fall back to UC-.  pat_init() updates these values to support
44  * more cache modes, WC and WT, when it is safe to do so.  See pat_init()
45  * for the details.  Note, __early_ioremap() used during early boot-time
46  * takes pgprot_t (pte encoding) and does not use these tables.
47  *
48  *   Index into __cachemode2pte_tbl[] is the cachemode.
49  *
50  *   Index into __pte2cachemode_tbl[] are the caching attribute bits of the pte
51  *   (_PAGE_PWT, _PAGE_PCD, _PAGE_PAT) at index bit positions 0, 1, 2.
52  */
53 static uint16_t __cachemode2pte_tbl[_PAGE_CACHE_MODE_NUM] = {
54 	[_PAGE_CACHE_MODE_WB      ]	= 0         | 0        ,
55 	[_PAGE_CACHE_MODE_WC      ]	= 0         | _PAGE_PCD,
56 	[_PAGE_CACHE_MODE_UC_MINUS]	= 0         | _PAGE_PCD,
57 	[_PAGE_CACHE_MODE_UC      ]	= _PAGE_PWT | _PAGE_PCD,
58 	[_PAGE_CACHE_MODE_WT      ]	= 0         | _PAGE_PCD,
59 	[_PAGE_CACHE_MODE_WP      ]	= 0         | _PAGE_PCD,
60 };
61 
62 unsigned long cachemode2protval(enum page_cache_mode pcm)
63 {
64 	if (likely(pcm == 0))
65 		return 0;
66 	return __cachemode2pte_tbl[pcm];
67 }
68 EXPORT_SYMBOL(cachemode2protval);
69 
70 static uint8_t __pte2cachemode_tbl[8] = {
71 	[__pte2cm_idx( 0        | 0         | 0        )] = _PAGE_CACHE_MODE_WB,
72 	[__pte2cm_idx(_PAGE_PWT | 0         | 0        )] = _PAGE_CACHE_MODE_UC_MINUS,
73 	[__pte2cm_idx( 0        | _PAGE_PCD | 0        )] = _PAGE_CACHE_MODE_UC_MINUS,
74 	[__pte2cm_idx(_PAGE_PWT | _PAGE_PCD | 0        )] = _PAGE_CACHE_MODE_UC,
75 	[__pte2cm_idx( 0        | 0         | _PAGE_PAT)] = _PAGE_CACHE_MODE_WB,
76 	[__pte2cm_idx(_PAGE_PWT | 0         | _PAGE_PAT)] = _PAGE_CACHE_MODE_UC_MINUS,
77 	[__pte2cm_idx(0         | _PAGE_PCD | _PAGE_PAT)] = _PAGE_CACHE_MODE_UC_MINUS,
78 	[__pte2cm_idx(_PAGE_PWT | _PAGE_PCD | _PAGE_PAT)] = _PAGE_CACHE_MODE_UC,
79 };
80 
81 /* Check that the write-protect PAT entry is set for write-protect */
82 bool x86_has_pat_wp(void)
83 {
84 	return __pte2cachemode_tbl[_PAGE_CACHE_MODE_WP] == _PAGE_CACHE_MODE_WP;
85 }
86 
87 enum page_cache_mode pgprot2cachemode(pgprot_t pgprot)
88 {
89 	unsigned long masked;
90 
91 	masked = pgprot_val(pgprot) & _PAGE_CACHE_MASK;
92 	if (likely(masked == 0))
93 		return 0;
94 	return __pte2cachemode_tbl[__pte2cm_idx(masked)];
95 }
96 
97 static unsigned long __initdata pgt_buf_start;
98 static unsigned long __initdata pgt_buf_end;
99 static unsigned long __initdata pgt_buf_top;
100 
101 static unsigned long min_pfn_mapped;
102 
103 static bool __initdata can_use_brk_pgt = true;
104 
105 /*
106  * Pages returned are already directly mapped.
107  *
108  * Changing that is likely to break Xen, see commit:
109  *
110  *    279b706 x86,xen: introduce x86_init.mapping.pagetable_reserve
111  *
112  * for detailed information.
113  */
114 __ref void *alloc_low_pages(unsigned int num)
115 {
116 	unsigned long pfn;
117 	int i;
118 
119 	if (after_bootmem) {
120 		unsigned int order;
121 
122 		order = get_order((unsigned long)num << PAGE_SHIFT);
123 		return (void *)__get_free_pages(GFP_ATOMIC | __GFP_ZERO, order);
124 	}
125 
126 	if ((pgt_buf_end + num) > pgt_buf_top || !can_use_brk_pgt) {
127 		unsigned long ret = 0;
128 
129 		if (min_pfn_mapped < max_pfn_mapped) {
130 			ret = memblock_find_in_range(
131 					min_pfn_mapped << PAGE_SHIFT,
132 					max_pfn_mapped << PAGE_SHIFT,
133 					PAGE_SIZE * num , PAGE_SIZE);
134 		}
135 		if (ret)
136 			memblock_reserve(ret, PAGE_SIZE * num);
137 		else if (can_use_brk_pgt)
138 			ret = __pa(extend_brk(PAGE_SIZE * num, PAGE_SIZE));
139 
140 		if (!ret)
141 			panic("alloc_low_pages: can not alloc memory");
142 
143 		pfn = ret >> PAGE_SHIFT;
144 	} else {
145 		pfn = pgt_buf_end;
146 		pgt_buf_end += num;
147 	}
148 
149 	for (i = 0; i < num; i++) {
150 		void *adr;
151 
152 		adr = __va((pfn + i) << PAGE_SHIFT);
153 		clear_page(adr);
154 	}
155 
156 	return __va(pfn << PAGE_SHIFT);
157 }
158 
159 /*
160  * By default need 3 4k for initial PMD_SIZE,  3 4k for 0-ISA_END_ADDRESS.
161  * With KASLR memory randomization, depending on the machine e820 memory
162  * and the PUD alignment. We may need twice more pages when KASLR memory
163  * randomization is enabled.
164  */
165 #ifndef CONFIG_RANDOMIZE_MEMORY
166 #define INIT_PGD_PAGE_COUNT      6
167 #else
168 #define INIT_PGD_PAGE_COUNT      12
169 #endif
170 #define INIT_PGT_BUF_SIZE	(INIT_PGD_PAGE_COUNT * PAGE_SIZE)
171 RESERVE_BRK(early_pgt_alloc, INIT_PGT_BUF_SIZE);
172 void  __init early_alloc_pgt_buf(void)
173 {
174 	unsigned long tables = INIT_PGT_BUF_SIZE;
175 	phys_addr_t base;
176 
177 	base = __pa(extend_brk(tables, PAGE_SIZE));
178 
179 	pgt_buf_start = base >> PAGE_SHIFT;
180 	pgt_buf_end = pgt_buf_start;
181 	pgt_buf_top = pgt_buf_start + (tables >> PAGE_SHIFT);
182 }
183 
184 int after_bootmem;
185 
186 early_param_on_off("gbpages", "nogbpages", direct_gbpages, CONFIG_X86_DIRECT_GBPAGES);
187 
188 struct map_range {
189 	unsigned long start;
190 	unsigned long end;
191 	unsigned page_size_mask;
192 };
193 
194 static int page_size_mask;
195 
196 /*
197  * Save some of cr4 feature set we're using (e.g.  Pentium 4MB
198  * enable and PPro Global page enable), so that any CPU's that boot
199  * up after us can get the correct flags. Invoked on the boot CPU.
200  */
201 static inline void cr4_set_bits_and_update_boot(unsigned long mask)
202 {
203 	mmu_cr4_features |= mask;
204 	if (trampoline_cr4_features)
205 		*trampoline_cr4_features = mmu_cr4_features;
206 	cr4_set_bits(mask);
207 }
208 
209 static void __init probe_page_size_mask(void)
210 {
211 	/*
212 	 * For pagealloc debugging, identity mapping will use small pages.
213 	 * This will simplify cpa(), which otherwise needs to support splitting
214 	 * large pages into small in interrupt context, etc.
215 	 */
216 	if (boot_cpu_has(X86_FEATURE_PSE) && !debug_pagealloc_enabled())
217 		page_size_mask |= 1 << PG_LEVEL_2M;
218 	else
219 		direct_gbpages = 0;
220 
221 	/* Enable PSE if available */
222 	if (boot_cpu_has(X86_FEATURE_PSE))
223 		cr4_set_bits_and_update_boot(X86_CR4_PSE);
224 
225 	/* Enable PGE if available */
226 	__supported_pte_mask &= ~_PAGE_GLOBAL;
227 	if (boot_cpu_has(X86_FEATURE_PGE)) {
228 		cr4_set_bits_and_update_boot(X86_CR4_PGE);
229 		__supported_pte_mask |= _PAGE_GLOBAL;
230 	}
231 
232 	/* By the default is everything supported: */
233 	__default_kernel_pte_mask = __supported_pte_mask;
234 	/* Except when with PTI where the kernel is mostly non-Global: */
235 	if (cpu_feature_enabled(X86_FEATURE_PTI))
236 		__default_kernel_pte_mask &= ~_PAGE_GLOBAL;
237 
238 	/* Enable 1 GB linear kernel mappings if available: */
239 	if (direct_gbpages && boot_cpu_has(X86_FEATURE_GBPAGES)) {
240 		printk(KERN_INFO "Using GB pages for direct mapping\n");
241 		page_size_mask |= 1 << PG_LEVEL_1G;
242 	} else {
243 		direct_gbpages = 0;
244 	}
245 }
246 
247 static void setup_pcid(void)
248 {
249 	if (!IS_ENABLED(CONFIG_X86_64))
250 		return;
251 
252 	if (!boot_cpu_has(X86_FEATURE_PCID))
253 		return;
254 
255 	if (boot_cpu_has(X86_FEATURE_PGE)) {
256 		/*
257 		 * This can't be cr4_set_bits_and_update_boot() -- the
258 		 * trampoline code can't handle CR4.PCIDE and it wouldn't
259 		 * do any good anyway.  Despite the name,
260 		 * cr4_set_bits_and_update_boot() doesn't actually cause
261 		 * the bits in question to remain set all the way through
262 		 * the secondary boot asm.
263 		 *
264 		 * Instead, we brute-force it and set CR4.PCIDE manually in
265 		 * start_secondary().
266 		 */
267 		cr4_set_bits(X86_CR4_PCIDE);
268 
269 		/*
270 		 * INVPCID's single-context modes (2/3) only work if we set
271 		 * X86_CR4_PCIDE, *and* we INVPCID support.  It's unusable
272 		 * on systems that have X86_CR4_PCIDE clear, or that have
273 		 * no INVPCID support at all.
274 		 */
275 		if (boot_cpu_has(X86_FEATURE_INVPCID))
276 			setup_force_cpu_cap(X86_FEATURE_INVPCID_SINGLE);
277 	} else {
278 		/*
279 		 * flush_tlb_all(), as currently implemented, won't work if
280 		 * PCID is on but PGE is not.  Since that combination
281 		 * doesn't exist on real hardware, there's no reason to try
282 		 * to fully support it, but it's polite to avoid corrupting
283 		 * data if we're on an improperly configured VM.
284 		 */
285 		setup_clear_cpu_cap(X86_FEATURE_PCID);
286 	}
287 }
288 
289 #ifdef CONFIG_X86_32
290 #define NR_RANGE_MR 3
291 #else /* CONFIG_X86_64 */
292 #define NR_RANGE_MR 5
293 #endif
294 
295 static int __meminit save_mr(struct map_range *mr, int nr_range,
296 			     unsigned long start_pfn, unsigned long end_pfn,
297 			     unsigned long page_size_mask)
298 {
299 	if (start_pfn < end_pfn) {
300 		if (nr_range >= NR_RANGE_MR)
301 			panic("run out of range for init_memory_mapping\n");
302 		mr[nr_range].start = start_pfn<<PAGE_SHIFT;
303 		mr[nr_range].end   = end_pfn<<PAGE_SHIFT;
304 		mr[nr_range].page_size_mask = page_size_mask;
305 		nr_range++;
306 	}
307 
308 	return nr_range;
309 }
310 
311 /*
312  * adjust the page_size_mask for small range to go with
313  *	big page size instead small one if nearby are ram too.
314  */
315 static void __ref adjust_range_page_size_mask(struct map_range *mr,
316 							 int nr_range)
317 {
318 	int i;
319 
320 	for (i = 0; i < nr_range; i++) {
321 		if ((page_size_mask & (1<<PG_LEVEL_2M)) &&
322 		    !(mr[i].page_size_mask & (1<<PG_LEVEL_2M))) {
323 			unsigned long start = round_down(mr[i].start, PMD_SIZE);
324 			unsigned long end = round_up(mr[i].end, PMD_SIZE);
325 
326 #ifdef CONFIG_X86_32
327 			if ((end >> PAGE_SHIFT) > max_low_pfn)
328 				continue;
329 #endif
330 
331 			if (memblock_is_region_memory(start, end - start))
332 				mr[i].page_size_mask |= 1<<PG_LEVEL_2M;
333 		}
334 		if ((page_size_mask & (1<<PG_LEVEL_1G)) &&
335 		    !(mr[i].page_size_mask & (1<<PG_LEVEL_1G))) {
336 			unsigned long start = round_down(mr[i].start, PUD_SIZE);
337 			unsigned long end = round_up(mr[i].end, PUD_SIZE);
338 
339 			if (memblock_is_region_memory(start, end - start))
340 				mr[i].page_size_mask |= 1<<PG_LEVEL_1G;
341 		}
342 	}
343 }
344 
345 static const char *page_size_string(struct map_range *mr)
346 {
347 	static const char str_1g[] = "1G";
348 	static const char str_2m[] = "2M";
349 	static const char str_4m[] = "4M";
350 	static const char str_4k[] = "4k";
351 
352 	if (mr->page_size_mask & (1<<PG_LEVEL_1G))
353 		return str_1g;
354 	/*
355 	 * 32-bit without PAE has a 4M large page size.
356 	 * PG_LEVEL_2M is misnamed, but we can at least
357 	 * print out the right size in the string.
358 	 */
359 	if (IS_ENABLED(CONFIG_X86_32) &&
360 	    !IS_ENABLED(CONFIG_X86_PAE) &&
361 	    mr->page_size_mask & (1<<PG_LEVEL_2M))
362 		return str_4m;
363 
364 	if (mr->page_size_mask & (1<<PG_LEVEL_2M))
365 		return str_2m;
366 
367 	return str_4k;
368 }
369 
370 static int __meminit split_mem_range(struct map_range *mr, int nr_range,
371 				     unsigned long start,
372 				     unsigned long end)
373 {
374 	unsigned long start_pfn, end_pfn, limit_pfn;
375 	unsigned long pfn;
376 	int i;
377 
378 	limit_pfn = PFN_DOWN(end);
379 
380 	/* head if not big page alignment ? */
381 	pfn = start_pfn = PFN_DOWN(start);
382 #ifdef CONFIG_X86_32
383 	/*
384 	 * Don't use a large page for the first 2/4MB of memory
385 	 * because there are often fixed size MTRRs in there
386 	 * and overlapping MTRRs into large pages can cause
387 	 * slowdowns.
388 	 */
389 	if (pfn == 0)
390 		end_pfn = PFN_DOWN(PMD_SIZE);
391 	else
392 		end_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
393 #else /* CONFIG_X86_64 */
394 	end_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
395 #endif
396 	if (end_pfn > limit_pfn)
397 		end_pfn = limit_pfn;
398 	if (start_pfn < end_pfn) {
399 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
400 		pfn = end_pfn;
401 	}
402 
403 	/* big page (2M) range */
404 	start_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
405 #ifdef CONFIG_X86_32
406 	end_pfn = round_down(limit_pfn, PFN_DOWN(PMD_SIZE));
407 #else /* CONFIG_X86_64 */
408 	end_pfn = round_up(pfn, PFN_DOWN(PUD_SIZE));
409 	if (end_pfn > round_down(limit_pfn, PFN_DOWN(PMD_SIZE)))
410 		end_pfn = round_down(limit_pfn, PFN_DOWN(PMD_SIZE));
411 #endif
412 
413 	if (start_pfn < end_pfn) {
414 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
415 				page_size_mask & (1<<PG_LEVEL_2M));
416 		pfn = end_pfn;
417 	}
418 
419 #ifdef CONFIG_X86_64
420 	/* big page (1G) range */
421 	start_pfn = round_up(pfn, PFN_DOWN(PUD_SIZE));
422 	end_pfn = round_down(limit_pfn, PFN_DOWN(PUD_SIZE));
423 	if (start_pfn < end_pfn) {
424 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
425 				page_size_mask &
426 				 ((1<<PG_LEVEL_2M)|(1<<PG_LEVEL_1G)));
427 		pfn = end_pfn;
428 	}
429 
430 	/* tail is not big page (1G) alignment */
431 	start_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
432 	end_pfn = round_down(limit_pfn, PFN_DOWN(PMD_SIZE));
433 	if (start_pfn < end_pfn) {
434 		nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
435 				page_size_mask & (1<<PG_LEVEL_2M));
436 		pfn = end_pfn;
437 	}
438 #endif
439 
440 	/* tail is not big page (2M) alignment */
441 	start_pfn = pfn;
442 	end_pfn = limit_pfn;
443 	nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
444 
445 	if (!after_bootmem)
446 		adjust_range_page_size_mask(mr, nr_range);
447 
448 	/* try to merge same page size and continuous */
449 	for (i = 0; nr_range > 1 && i < nr_range - 1; i++) {
450 		unsigned long old_start;
451 		if (mr[i].end != mr[i+1].start ||
452 		    mr[i].page_size_mask != mr[i+1].page_size_mask)
453 			continue;
454 		/* move it */
455 		old_start = mr[i].start;
456 		memmove(&mr[i], &mr[i+1],
457 			(nr_range - 1 - i) * sizeof(struct map_range));
458 		mr[i--].start = old_start;
459 		nr_range--;
460 	}
461 
462 	for (i = 0; i < nr_range; i++)
463 		pr_debug(" [mem %#010lx-%#010lx] page %s\n",
464 				mr[i].start, mr[i].end - 1,
465 				page_size_string(&mr[i]));
466 
467 	return nr_range;
468 }
469 
470 struct range pfn_mapped[E820_MAX_ENTRIES];
471 int nr_pfn_mapped;
472 
473 static void add_pfn_range_mapped(unsigned long start_pfn, unsigned long end_pfn)
474 {
475 	nr_pfn_mapped = add_range_with_merge(pfn_mapped, E820_MAX_ENTRIES,
476 					     nr_pfn_mapped, start_pfn, end_pfn);
477 	nr_pfn_mapped = clean_sort_range(pfn_mapped, E820_MAX_ENTRIES);
478 
479 	max_pfn_mapped = max(max_pfn_mapped, end_pfn);
480 
481 	if (start_pfn < (1UL<<(32-PAGE_SHIFT)))
482 		max_low_pfn_mapped = max(max_low_pfn_mapped,
483 					 min(end_pfn, 1UL<<(32-PAGE_SHIFT)));
484 }
485 
486 bool pfn_range_is_mapped(unsigned long start_pfn, unsigned long end_pfn)
487 {
488 	int i;
489 
490 	for (i = 0; i < nr_pfn_mapped; i++)
491 		if ((start_pfn >= pfn_mapped[i].start) &&
492 		    (end_pfn <= pfn_mapped[i].end))
493 			return true;
494 
495 	return false;
496 }
497 
498 /*
499  * Setup the direct mapping of the physical memory at PAGE_OFFSET.
500  * This runs before bootmem is initialized and gets pages directly from
501  * the physical memory. To access them they are temporarily mapped.
502  */
503 unsigned long __ref init_memory_mapping(unsigned long start,
504 					unsigned long end, pgprot_t prot)
505 {
506 	struct map_range mr[NR_RANGE_MR];
507 	unsigned long ret = 0;
508 	int nr_range, i;
509 
510 	pr_debug("init_memory_mapping: [mem %#010lx-%#010lx]\n",
511 	       start, end - 1);
512 
513 	memset(mr, 0, sizeof(mr));
514 	nr_range = split_mem_range(mr, 0, start, end);
515 
516 	for (i = 0; i < nr_range; i++)
517 		ret = kernel_physical_mapping_init(mr[i].start, mr[i].end,
518 						   mr[i].page_size_mask,
519 						   prot);
520 
521 	add_pfn_range_mapped(start >> PAGE_SHIFT, ret >> PAGE_SHIFT);
522 
523 	return ret >> PAGE_SHIFT;
524 }
525 
526 /*
527  * We need to iterate through the E820 memory map and create direct mappings
528  * for only E820_TYPE_RAM and E820_KERN_RESERVED regions. We cannot simply
529  * create direct mappings for all pfns from [0 to max_low_pfn) and
530  * [4GB to max_pfn) because of possible memory holes in high addresses
531  * that cannot be marked as UC by fixed/variable range MTRRs.
532  * Depending on the alignment of E820 ranges, this may possibly result
533  * in using smaller size (i.e. 4K instead of 2M or 1G) page tables.
534  *
535  * init_mem_mapping() calls init_range_memory_mapping() with big range.
536  * That range would have hole in the middle or ends, and only ram parts
537  * will be mapped in init_range_memory_mapping().
538  */
539 static unsigned long __init init_range_memory_mapping(
540 					   unsigned long r_start,
541 					   unsigned long r_end)
542 {
543 	unsigned long start_pfn, end_pfn;
544 	unsigned long mapped_ram_size = 0;
545 	int i;
546 
547 	for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, NULL) {
548 		u64 start = clamp_val(PFN_PHYS(start_pfn), r_start, r_end);
549 		u64 end = clamp_val(PFN_PHYS(end_pfn), r_start, r_end);
550 		if (start >= end)
551 			continue;
552 
553 		/*
554 		 * if it is overlapping with brk pgt, we need to
555 		 * alloc pgt buf from memblock instead.
556 		 */
557 		can_use_brk_pgt = max(start, (u64)pgt_buf_end<<PAGE_SHIFT) >=
558 				    min(end, (u64)pgt_buf_top<<PAGE_SHIFT);
559 		init_memory_mapping(start, end, PAGE_KERNEL);
560 		mapped_ram_size += end - start;
561 		can_use_brk_pgt = true;
562 	}
563 
564 	return mapped_ram_size;
565 }
566 
567 static unsigned long __init get_new_step_size(unsigned long step_size)
568 {
569 	/*
570 	 * Initial mapped size is PMD_SIZE (2M).
571 	 * We can not set step_size to be PUD_SIZE (1G) yet.
572 	 * In worse case, when we cross the 1G boundary, and
573 	 * PG_LEVEL_2M is not set, we will need 1+1+512 pages (2M + 8k)
574 	 * to map 1G range with PTE. Hence we use one less than the
575 	 * difference of page table level shifts.
576 	 *
577 	 * Don't need to worry about overflow in the top-down case, on 32bit,
578 	 * when step_size is 0, round_down() returns 0 for start, and that
579 	 * turns it into 0x100000000ULL.
580 	 * In the bottom-up case, round_up(x, 0) returns 0 though too, which
581 	 * needs to be taken into consideration by the code below.
582 	 */
583 	return step_size << (PMD_SHIFT - PAGE_SHIFT - 1);
584 }
585 
586 /**
587  * memory_map_top_down - Map [map_start, map_end) top down
588  * @map_start: start address of the target memory range
589  * @map_end: end address of the target memory range
590  *
591  * This function will setup direct mapping for memory range
592  * [map_start, map_end) in top-down. That said, the page tables
593  * will be allocated at the end of the memory, and we map the
594  * memory in top-down.
595  */
596 static void __init memory_map_top_down(unsigned long map_start,
597 				       unsigned long map_end)
598 {
599 	unsigned long real_end, last_start;
600 	unsigned long step_size;
601 	unsigned long addr;
602 	unsigned long mapped_ram_size = 0;
603 
604 	/* xen has big range in reserved near end of ram, skip it at first.*/
605 	addr = memblock_find_in_range(map_start, map_end, PMD_SIZE, PMD_SIZE);
606 	real_end = addr + PMD_SIZE;
607 
608 	/* step_size need to be small so pgt_buf from BRK could cover it */
609 	step_size = PMD_SIZE;
610 	max_pfn_mapped = 0; /* will get exact value next */
611 	min_pfn_mapped = real_end >> PAGE_SHIFT;
612 	last_start = real_end;
613 
614 	/*
615 	 * We start from the top (end of memory) and go to the bottom.
616 	 * The memblock_find_in_range() gets us a block of RAM from the
617 	 * end of RAM in [min_pfn_mapped, max_pfn_mapped) used as new pages
618 	 * for page table.
619 	 */
620 	while (last_start > map_start) {
621 		unsigned long start;
622 
623 		if (last_start > step_size) {
624 			start = round_down(last_start - 1, step_size);
625 			if (start < map_start)
626 				start = map_start;
627 		} else
628 			start = map_start;
629 		mapped_ram_size += init_range_memory_mapping(start,
630 							last_start);
631 		last_start = start;
632 		min_pfn_mapped = last_start >> PAGE_SHIFT;
633 		if (mapped_ram_size >= step_size)
634 			step_size = get_new_step_size(step_size);
635 	}
636 
637 	if (real_end < map_end)
638 		init_range_memory_mapping(real_end, map_end);
639 }
640 
641 /**
642  * memory_map_bottom_up - Map [map_start, map_end) bottom up
643  * @map_start: start address of the target memory range
644  * @map_end: end address of the target memory range
645  *
646  * This function will setup direct mapping for memory range
647  * [map_start, map_end) in bottom-up. Since we have limited the
648  * bottom-up allocation above the kernel, the page tables will
649  * be allocated just above the kernel and we map the memory
650  * in [map_start, map_end) in bottom-up.
651  */
652 static void __init memory_map_bottom_up(unsigned long map_start,
653 					unsigned long map_end)
654 {
655 	unsigned long next, start;
656 	unsigned long mapped_ram_size = 0;
657 	/* step_size need to be small so pgt_buf from BRK could cover it */
658 	unsigned long step_size = PMD_SIZE;
659 
660 	start = map_start;
661 	min_pfn_mapped = start >> PAGE_SHIFT;
662 
663 	/*
664 	 * We start from the bottom (@map_start) and go to the top (@map_end).
665 	 * The memblock_find_in_range() gets us a block of RAM from the
666 	 * end of RAM in [min_pfn_mapped, max_pfn_mapped) used as new pages
667 	 * for page table.
668 	 */
669 	while (start < map_end) {
670 		if (step_size && map_end - start > step_size) {
671 			next = round_up(start + 1, step_size);
672 			if (next > map_end)
673 				next = map_end;
674 		} else {
675 			next = map_end;
676 		}
677 
678 		mapped_ram_size += init_range_memory_mapping(start, next);
679 		start = next;
680 
681 		if (mapped_ram_size >= step_size)
682 			step_size = get_new_step_size(step_size);
683 	}
684 }
685 
686 /*
687  * The real mode trampoline, which is required for bootstrapping CPUs
688  * occupies only a small area under the low 1MB.  See reserve_real_mode()
689  * for details.
690  *
691  * If KASLR is disabled the first PGD entry of the direct mapping is copied
692  * to map the real mode trampoline.
693  *
694  * If KASLR is enabled, copy only the PUD which covers the low 1MB
695  * area. This limits the randomization granularity to 1GB for both 4-level
696  * and 5-level paging.
697  */
698 static void __init init_trampoline(void)
699 {
700 #ifdef CONFIG_X86_64
701 	if (!kaslr_memory_enabled())
702 		trampoline_pgd_entry = init_top_pgt[pgd_index(__PAGE_OFFSET)];
703 	else
704 		init_trampoline_kaslr();
705 #endif
706 }
707 
708 void __init init_mem_mapping(void)
709 {
710 	unsigned long end;
711 
712 	pti_check_boottime_disable();
713 	probe_page_size_mask();
714 	setup_pcid();
715 
716 #ifdef CONFIG_X86_64
717 	end = max_pfn << PAGE_SHIFT;
718 #else
719 	end = max_low_pfn << PAGE_SHIFT;
720 #endif
721 
722 	/* the ISA range is always mapped regardless of memory holes */
723 	init_memory_mapping(0, ISA_END_ADDRESS, PAGE_KERNEL);
724 
725 	/* Init the trampoline, possibly with KASLR memory offset */
726 	init_trampoline();
727 
728 	/*
729 	 * If the allocation is in bottom-up direction, we setup direct mapping
730 	 * in bottom-up, otherwise we setup direct mapping in top-down.
731 	 */
732 	if (memblock_bottom_up()) {
733 		unsigned long kernel_end = __pa_symbol(_end);
734 
735 		/*
736 		 * we need two separate calls here. This is because we want to
737 		 * allocate page tables above the kernel. So we first map
738 		 * [kernel_end, end) to make memory above the kernel be mapped
739 		 * as soon as possible. And then use page tables allocated above
740 		 * the kernel to map [ISA_END_ADDRESS, kernel_end).
741 		 */
742 		memory_map_bottom_up(kernel_end, end);
743 		memory_map_bottom_up(ISA_END_ADDRESS, kernel_end);
744 	} else {
745 		memory_map_top_down(ISA_END_ADDRESS, end);
746 	}
747 
748 #ifdef CONFIG_X86_64
749 	if (max_pfn > max_low_pfn) {
750 		/* can we preseve max_low_pfn ?*/
751 		max_low_pfn = max_pfn;
752 	}
753 #else
754 	early_ioremap_page_table_range_init();
755 #endif
756 
757 	load_cr3(swapper_pg_dir);
758 	__flush_tlb_all();
759 
760 	x86_init.hyper.init_mem_mapping();
761 
762 	early_memtest(0, max_pfn_mapped << PAGE_SHIFT);
763 }
764 
765 /*
766  * Initialize an mm_struct to be used during poking and a pointer to be used
767  * during patching.
768  */
769 void __init poking_init(void)
770 {
771 	spinlock_t *ptl;
772 	pte_t *ptep;
773 
774 	poking_mm = copy_init_mm();
775 	BUG_ON(!poking_mm);
776 
777 	/*
778 	 * Randomize the poking address, but make sure that the following page
779 	 * will be mapped at the same PMD. We need 2 pages, so find space for 3,
780 	 * and adjust the address if the PMD ends after the first one.
781 	 */
782 	poking_addr = TASK_UNMAPPED_BASE;
783 	if (IS_ENABLED(CONFIG_RANDOMIZE_BASE))
784 		poking_addr += (kaslr_get_random_long("Poking") & PAGE_MASK) %
785 			(TASK_SIZE - TASK_UNMAPPED_BASE - 3 * PAGE_SIZE);
786 
787 	if (((poking_addr + PAGE_SIZE) & ~PMD_MASK) == 0)
788 		poking_addr += PAGE_SIZE;
789 
790 	/*
791 	 * We need to trigger the allocation of the page-tables that will be
792 	 * needed for poking now. Later, poking may be performed in an atomic
793 	 * section, which might cause allocation to fail.
794 	 */
795 	ptep = get_locked_pte(poking_mm, poking_addr, &ptl);
796 	BUG_ON(!ptep);
797 	pte_unmap_unlock(ptep, ptl);
798 }
799 
800 /*
801  * devmem_is_allowed() checks to see if /dev/mem access to a certain address
802  * is valid. The argument is a physical page number.
803  *
804  * On x86, access has to be given to the first megabyte of RAM because that
805  * area traditionally contains BIOS code and data regions used by X, dosemu,
806  * and similar apps. Since they map the entire memory range, the whole range
807  * must be allowed (for mapping), but any areas that would otherwise be
808  * disallowed are flagged as being "zero filled" instead of rejected.
809  * Access has to be given to non-kernel-ram areas as well, these contain the
810  * PCI mmio resources as well as potential bios/acpi data regions.
811  */
812 int devmem_is_allowed(unsigned long pagenr)
813 {
814 	if (region_intersects(PFN_PHYS(pagenr), PAGE_SIZE,
815 				IORESOURCE_SYSTEM_RAM, IORES_DESC_NONE)
816 			!= REGION_DISJOINT) {
817 		/*
818 		 * For disallowed memory regions in the low 1MB range,
819 		 * request that the page be shown as all zeros.
820 		 */
821 		if (pagenr < 256)
822 			return 2;
823 
824 		return 0;
825 	}
826 
827 	/*
828 	 * This must follow RAM test, since System RAM is considered a
829 	 * restricted resource under CONFIG_STRICT_IOMEM.
830 	 */
831 	if (iomem_is_exclusive(pagenr << PAGE_SHIFT)) {
832 		/* Low 1MB bypasses iomem restrictions. */
833 		if (pagenr < 256)
834 			return 1;
835 
836 		return 0;
837 	}
838 
839 	return 1;
840 }
841 
842 void free_init_pages(const char *what, unsigned long begin, unsigned long end)
843 {
844 	unsigned long begin_aligned, end_aligned;
845 
846 	/* Make sure boundaries are page aligned */
847 	begin_aligned = PAGE_ALIGN(begin);
848 	end_aligned   = end & PAGE_MASK;
849 
850 	if (WARN_ON(begin_aligned != begin || end_aligned != end)) {
851 		begin = begin_aligned;
852 		end   = end_aligned;
853 	}
854 
855 	if (begin >= end)
856 		return;
857 
858 	/*
859 	 * If debugging page accesses then do not free this memory but
860 	 * mark them not present - any buggy init-section access will
861 	 * create a kernel page fault:
862 	 */
863 	if (debug_pagealloc_enabled()) {
864 		pr_info("debug: unmapping init [mem %#010lx-%#010lx]\n",
865 			begin, end - 1);
866 		/*
867 		 * Inform kmemleak about the hole in the memory since the
868 		 * corresponding pages will be unmapped.
869 		 */
870 		kmemleak_free_part((void *)begin, end - begin);
871 		set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
872 	} else {
873 		/*
874 		 * We just marked the kernel text read only above, now that
875 		 * we are going to free part of that, we need to make that
876 		 * writeable and non-executable first.
877 		 */
878 		set_memory_nx(begin, (end - begin) >> PAGE_SHIFT);
879 		set_memory_rw(begin, (end - begin) >> PAGE_SHIFT);
880 
881 		free_reserved_area((void *)begin, (void *)end,
882 				   POISON_FREE_INITMEM, what);
883 	}
884 }
885 
886 /*
887  * begin/end can be in the direct map or the "high kernel mapping"
888  * used for the kernel image only.  free_init_pages() will do the
889  * right thing for either kind of address.
890  */
891 void free_kernel_image_pages(const char *what, void *begin, void *end)
892 {
893 	unsigned long begin_ul = (unsigned long)begin;
894 	unsigned long end_ul = (unsigned long)end;
895 	unsigned long len_pages = (end_ul - begin_ul) >> PAGE_SHIFT;
896 
897 	free_init_pages(what, begin_ul, end_ul);
898 
899 	/*
900 	 * PTI maps some of the kernel into userspace.  For performance,
901 	 * this includes some kernel areas that do not contain secrets.
902 	 * Those areas might be adjacent to the parts of the kernel image
903 	 * being freed, which may contain secrets.  Remove the "high kernel
904 	 * image mapping" for these freed areas, ensuring they are not even
905 	 * potentially vulnerable to Meltdown regardless of the specific
906 	 * optimizations PTI is currently using.
907 	 *
908 	 * The "noalias" prevents unmapping the direct map alias which is
909 	 * needed to access the freed pages.
910 	 *
911 	 * This is only valid for 64bit kernels. 32bit has only one mapping
912 	 * which can't be treated in this way for obvious reasons.
913 	 */
914 	if (IS_ENABLED(CONFIG_X86_64) && cpu_feature_enabled(X86_FEATURE_PTI))
915 		set_memory_np_noalias(begin_ul, len_pages);
916 }
917 
918 void __ref free_initmem(void)
919 {
920 	e820__reallocate_tables();
921 
922 	mem_encrypt_free_decrypted_mem();
923 
924 	free_kernel_image_pages("unused kernel image (initmem)",
925 				&__init_begin, &__init_end);
926 }
927 
928 #ifdef CONFIG_BLK_DEV_INITRD
929 void __init free_initrd_mem(unsigned long start, unsigned long end)
930 {
931 	/*
932 	 * end could be not aligned, and We can not align that,
933 	 * decompresser could be confused by aligned initrd_end
934 	 * We already reserve the end partial page before in
935 	 *   - i386_start_kernel()
936 	 *   - x86_64_start_kernel()
937 	 *   - relocate_initrd()
938 	 * So here We can do PAGE_ALIGN() safely to get partial page to be freed
939 	 */
940 	free_init_pages("initrd", start, PAGE_ALIGN(end));
941 }
942 #endif
943 
944 /*
945  * Calculate the precise size of the DMA zone (first 16 MB of RAM),
946  * and pass it to the MM layer - to help it set zone watermarks more
947  * accurately.
948  *
949  * Done on 64-bit systems only for the time being, although 32-bit systems
950  * might benefit from this as well.
951  */
952 void __init memblock_find_dma_reserve(void)
953 {
954 #ifdef CONFIG_X86_64
955 	u64 nr_pages = 0, nr_free_pages = 0;
956 	unsigned long start_pfn, end_pfn;
957 	phys_addr_t start_addr, end_addr;
958 	int i;
959 	u64 u;
960 
961 	/*
962 	 * Iterate over all memory ranges (free and reserved ones alike),
963 	 * to calculate the total number of pages in the first 16 MB of RAM:
964 	 */
965 	nr_pages = 0;
966 	for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, NULL) {
967 		start_pfn = min(start_pfn, MAX_DMA_PFN);
968 		end_pfn   = min(end_pfn,   MAX_DMA_PFN);
969 
970 		nr_pages += end_pfn - start_pfn;
971 	}
972 
973 	/*
974 	 * Iterate over free memory ranges to calculate the number of free
975 	 * pages in the DMA zone, while not counting potential partial
976 	 * pages at the beginning or the end of the range:
977 	 */
978 	nr_free_pages = 0;
979 	for_each_free_mem_range(u, NUMA_NO_NODE, MEMBLOCK_NONE, &start_addr, &end_addr, NULL) {
980 		start_pfn = min_t(unsigned long, PFN_UP(start_addr), MAX_DMA_PFN);
981 		end_pfn   = min_t(unsigned long, PFN_DOWN(end_addr), MAX_DMA_PFN);
982 
983 		if (start_pfn < end_pfn)
984 			nr_free_pages += end_pfn - start_pfn;
985 	}
986 
987 	set_dma_reserve(nr_pages - nr_free_pages);
988 #endif
989 }
990 
991 void __init zone_sizes_init(void)
992 {
993 	unsigned long max_zone_pfns[MAX_NR_ZONES];
994 
995 	memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
996 
997 #ifdef CONFIG_ZONE_DMA
998 	max_zone_pfns[ZONE_DMA]		= min(MAX_DMA_PFN, max_low_pfn);
999 #endif
1000 #ifdef CONFIG_ZONE_DMA32
1001 	max_zone_pfns[ZONE_DMA32]	= min(MAX_DMA32_PFN, max_low_pfn);
1002 #endif
1003 	max_zone_pfns[ZONE_NORMAL]	= max_low_pfn;
1004 #ifdef CONFIG_HIGHMEM
1005 	max_zone_pfns[ZONE_HIGHMEM]	= max_pfn;
1006 #endif
1007 
1008 	free_area_init(max_zone_pfns);
1009 }
1010 
1011 __visible DEFINE_PER_CPU_SHARED_ALIGNED(struct tlb_state, cpu_tlbstate) = {
1012 	.loaded_mm = &init_mm,
1013 	.next_asid = 1,
1014 	.cr4 = ~0UL,	/* fail hard if we screw up cr4 shadow initialization */
1015 };
1016 
1017 void update_cache_mode_entry(unsigned entry, enum page_cache_mode cache)
1018 {
1019 	/* entry 0 MUST be WB (hardwired to speed up translations) */
1020 	BUG_ON(!entry && cache != _PAGE_CACHE_MODE_WB);
1021 
1022 	__cachemode2pte_tbl[cache] = __cm_idx2pte(entry);
1023 	__pte2cachemode_tbl[entry] = cache;
1024 }
1025 
1026 #ifdef CONFIG_SWAP
1027 unsigned long max_swapfile_size(void)
1028 {
1029 	unsigned long pages;
1030 
1031 	pages = generic_max_swapfile_size();
1032 
1033 	if (boot_cpu_has_bug(X86_BUG_L1TF) && l1tf_mitigation != L1TF_MITIGATION_OFF) {
1034 		/* Limit the swap file size to MAX_PA/2 for L1TF workaround */
1035 		unsigned long long l1tf_limit = l1tf_pfn_limit();
1036 		/*
1037 		 * We encode swap offsets also with 3 bits below those for pfn
1038 		 * which makes the usable limit higher.
1039 		 */
1040 #if CONFIG_PGTABLE_LEVELS > 2
1041 		l1tf_limit <<= PAGE_SHIFT - SWP_OFFSET_FIRST_BIT;
1042 #endif
1043 		pages = min_t(unsigned long long, l1tf_limit, pages);
1044 	}
1045 	return pages;
1046 }
1047 #endif
1048